it's an AMD GPU and besides this software tech is vendor agnostic so it's should theoretically work on any card even as far as 1080ti's
The Verge: AMD's DLSS competitor, FSR 2.0, is coming to Xbox
- Thread starter digitalrelic
- Start date
it's an AMD GPU and besides this software tech is vendor agnostic so it's should theoretically work on any card even as far as 1080ti's
It was my understanding that doing both simultaneously is costly. As resources are spent to achieve either method. I know that both can be done together, there isn't a technical reason they can't. I was just of the impression that its usually one or the other. Or primarily one and the otherused as a fail-safe if both are used in an engine.
Thanks for the correction though.
I'm hoping this leads to even better RT implementation with performance and resolution intact!
This is exactly what I hoped for!
The new consoles already are amazing as is, this will benefit them even more. I really don't see the need for midgen upgrades now (I barely did before).
I'm not the biggest fan of the console but this will also do wonders for the XSS' general resolution and RT implementation, I imagine.
This is exactly what I hoped for!
The new consoles already are amazing as is, this will benefit them even more. I really don't see the need for midgen upgrades now (I barely did before).
I'm not the biggest fan of the console but this will also do wonders for the XSS' general resolution and RT implementation, I imagine.
Certain bc it requires implementation for the motion vector aspect. Fsr 1.0 is now called RSR and is driver level on pc and can be truly applied to any game with no developer necessary.
This is us getting ahead of ourselves and not taking things into context.
As is, the new consoles are also adopting various rez gimmicks. Be that DRS or upsampling. This is just one more upsampling method to add to a long list of currently employed methods. This is not doing anything that is not already being done in some shape or form. At the end of the day, this is still just reconstruction.
Here are the facts. Running at native 1440p and outputting at that resolution, has a render cost less than running internally at 1440p and upsampling to 4k using any upsampling method. You add the cost of upsampling on top of whatever native rez render cost is being used. Where this tech can shine, or tech like these is if allowing you to render internally at say 1080p or 1252p and then this samples it up to 4k convincingly. But at a render cost that is say on par with you just running natively at 1440p. So your render cost s equivalent t 1440p, but your IQ is equivalent to say 1800p or 2000p.
But none of this means anything until we can see just what it costs. If it costs too much, then devs would opt for cheaper methods. That is what we should be looking at.
The high level description AMD gave of FSR 2 is indistinguishable from the description of various temporal up-sampling techniques going back several years now. Whether it's useful or not will basically depend on whether it's better in practice conceptually similar solutions available on other engines. For UE5, it's got an uphill battle against TSR, which has very good quality compared to last gen solutions.
The more useful case might be games running on various custom engines (a shrinking pool, to be sure) but which either do not have temporal up-sampling, or have underwhelming versions thereof. For UE5 games, of which there will be zillions in the coming years (plus UE4 games with TSR back ported like Ghostwire), it doesn't seem like it'll be useful unless the quality is of particular note.
FSR 1, the older version, can be hacked in to any game in theory (but then also applies upscaling to the HUD which is slightly suboptimal). Steam Deck has that option for example.
FSR 2, the new one, will not be able to do the same, because it requires information that is only available from within the game engine and which not every game may produce at all. So it will require support from the developers on a per-game basis.
The more useful case might be games running on various custom engines (a shrinking pool, to be sure) but which either do not have temporal up-sampling, or have underwhelming versions thereof. For UE5 games, of which there will be zillions in the coming years (plus UE4 games with TSR back ported like Ghostwire), it doesn't seem like it'll be useful unless the quality is of particular note.
FSR 1, the older version, can be hacked in to any game in theory (but then also applies upscaling to the HUD which is slightly suboptimal). Steam Deck has that option for example.
FSR 2, the new one, will not be able to do the same, because it requires information that is only available from within the game engine and which not every game may produce at all. So it will require support from the developers on a per-game basis.
Because Sony is slow to adopt new shit. Take VRR for example.
The Coalition is an Unreal studio and will very highly likely just use TSR as they have demoed with Alpha Point.
Which DLSS does.
This seems to be getting us similar results. So no, I'm not getting ahead of myself - Again, if it's as good as we're all hoping that it is.
I don't know why you felt the need to mention that other devs already use reconstruction techniques... Because... I never suggested they didn't?
This just seems to finally rival what Nvidia is doing with DLSS. Let's hope it does!
2.0 seems better than 1.0 from what I've quickly researched which is nice to see. Really hope tech like this makes it's way to consoles quickly. If the tools built into the SDK I don't see why it would be hard to implement. Especially in performance modes where it might be a 1440p image. DLSS is pretty damn magical, hopefully that switch rumor is true with that.
FSR 1.0 was also marketed only with Xbox in mind.
First game with FSR 1.0 was on PS5 ;)
Sure.
Thats why first game with FSR 1.0 which was also marketed with xbox, not PlayStation came first on PS5.
VRR is overdue, but cut those bs…
OP
OP
We're just going off of what AMD is claiming right now, which is that it achieves similar results & performance impact as DLSS 2.0+. If that's the case, this is going to be a really big leap forward on consoles.
No, not necessarily. Unreal engine games can use whatever they want but it's likely tsr will be superior and it's baked in, so…
OP
OP
My god dude, give it up lol. You don't need to reiterate the same thing 3 times in 3 separate posts in the same thread over the span of 2 minutes. We get it.
Some engines (like insomniacs one) have already temporal upscaling solution. FSR is for those who have not their own propietary tech
TAA and temporal reconstruction are not very different. In fact at the back end they're usually the same thing, the difference being whether the input resolution is the same as the output resolution or not. Specific algorithms will differ, but specific TAA algrithms and specific TAAu algorithms also differ from each other quite a lot in the finer details, so please understand I'm speaking on a fairly high level here.
1440p -> 4k temporal reconstruction vs 4k native TAA differs on how many samples they take per frame, but in both cases they accumulate samples across several frames to produce very high resolution images, which are then down-sampled to the final display. They might accumulate samples across 4, 8 or 16 frames, which implies an extremely large sample count, internally creating an image that may be somewhere in the region of 4x or 8x supersampled, which is then down-scaled to the target output 4k display. The anti-aliasing effect of TAA is produced by this intermediate step where you create this very high resolution image. But even if you started at only 1440p equivalent samples per frame, the intermediate image before output to the screen is still much higher res than it would have been rendering at native res without any form of AA - so it's still an "Anti aliased image".
Note here, that early work on TAA referred to the technique as "temporally-ammortized supersampling". I hope based on the above description that makes sense why they would have called it that, and how temporal upsampling / reconstruction is, conceptually, just a slight variation on TAA itself.
Now, regarding AA, yes they might also other techniques on top of this temporal stuff at some stage of the pipeline. It's not uncommon to use cheap post-process AA like FXAA at some stage of the image, to further mitigate some kinds of aliasing beyond that which the TAA suppresses, and (critically) to provide "good enough" anti aliasing coverage such that things which disrupt the TAA don't result in a momentarily completely raw image. Examples of things like that would be camera cuts - the algorithm needs to throw out the accumulated data when a camera cut occurs because the data of the old image has no bearing on the new one and if they kept it, the image would become smoother but dreadfully smeared / ghosted with the old image. Thus, in rejecting the data to prevent that, you wind up with 1 frame with no AA at all otherwise, which then gradually builds up as each new frame is generated. The FXAA isn't as good as TAA, but when it only needs to cover 2-3 frames before the TAA can start looking good again, that's usually not perceptually noticeable for users. A specific concrete example of games that do this is Resident Evil 2 / 3 Remake, but there are many more than that. The CoD games have been using a technique they refer to as "Flimic SMAA" which blends the SMAA algorithm with temporal accumulation to achieve their total solution. Sometimes in a menu you'll see a CryEngine game have options like "SMAA T2X" or whatever which is also similar.
I am not knocking the tech, or what it may or may not mean. I am just slightly skeptical and simply waiting for more information before we start hoping this up to be the next best thing for gaming. Again, we should see what it actually costs to implement, and what kinda results those costs give us.
For reference, DLSS on a 2070s (closest approximation to both a PS5/XSX) has a render cost of ~2.1ms for a 4K output. And that s with dedicated ML hardware.
OP
OP
From The Verge article, this is what AMD claims:
The Verge said:
Quality Mode:
Performance mode:
Last edited:
Note that even dlss isn't able to really match the native iq at a 4x (1080p -> 4k) upscale, so hopefully people are keeping their expectations in check
I really love the way DLSS takes care of all the shimmering artifacts and pixel crawling, to the point I sometimes prefer DLSS to native.
I'd happily sacrifice some sharpness if that means cleaning up the noise, not to mention the huge performance gains.
If this is as good as they claim it can't come soon enough to consoles.
I'd happily sacrifice some sharpness if that means cleaning up the noise, not to mention the huge performance gains.
If this is as good as they claim it can't come soon enough to consoles.
Seems like a potential game changer for game consoles and getting extra juice out of these boxes.
OP
OP
Going to be interesting to see if Deathloop includes FSR 2.0 when it launches on Xbox later this year.
In what context? I highly doubt Alex said that FSR 2.0 is more like CBR.
That was an amazing explanation, thanks for putting in the time!
Let me have a look at it again.
Edit: disregard, he said it's basically a version TAAU. That's what happens when you're lying in bed with COVID and your head is spinning 🤒
AMD GPUOpen said:
This specific quote makes me wonder if AMD is implying that there can be some hand optimization by the developer on a game by game basis to help cover any image quality gap that ML might have?
OP
OP
Kind of sounds like it. Maybe able to manually tweak the Algorithm on a scene-by-scene basis if needed?
This has nothing to do with Sony, it's an open standard that anyone can use. And regardless, Sony is not used to quickly publically announce what's in or is not in their console SDK.
Situation is same as with FSR 1.0. If devs want it, they can add it.
Given how most games just use the default dlss sharpening options, I'm a bit skeptical that many devs would spend that much time hyper optimizing their settings on a power scene basis
Exactly.
No one knew the PS5 could use FSR 1.0....until it was spotted in a game running on PS5, and devs mentioned it in articles.
Some things Sony dont even publicly talk about wrt the PS5. So far at least.